Make TS adapter spawning degrade gracefully instead of silently or fatally

[alongd]

Two failure modes in Scheduler.spawn_ts_jobs, both found on the same campaign run, both costing work that had already succeeded.

1. No eligible adapter — silently

An adapter is spawned only if it is in both the configured ts_adapters and ts_adapters_by_rmg_family[rxn.family]. When those lists do not intersect, the loop runs to completion, spawns nothing, and logs nothing.

Nothing recovers from it:

• tsg_spawned is latched True before the loop, so the reaction is never revisited;
• admit_unknown_family requires not family_known, so a known family with no eligible adapter cannot fall back to ts_adapters_for_unknown_unimolecular;
• the TS guess report is written with successful_methods: [] and unsuccessful_methods: [] — both empty, because nothing was attempted;
• the only user-visible symptom is TS did not converge, hours later, indistinguishable from having tried every adapter and failed.

Hit when ~/.arc/settings.py pinned:

ts_adapters = ['heuristics', 'AutoTST', 'GCN', 'xtb_gsm', 'orca_neb']

predating linear joining ARC's own default in arc/settings/settings.py:114. The family 1,2_Insertion_CO admits only ['kinbot', 'goflow', 'rits', 'linear']:

EFFECTIVE ts_adapters (after ~/.arc override): ['heuristics', 'autotst', 'gcn', 'xtb_gsm', 'orca_neb']
eligible for 1,2_Insertion_CO:                 ['kinbot', 'goflow', 'rits', 'linear']
--> adapters that spawn: []   (count=0)

reproducing the observed report exactly:

TS0:
  family: 1,2_Insertion_CO
  successful_methods: []
  unsuccessful_methods: []
  ts_guesses: {}

Worth naming the general hazard: arc/imports.py does settings.update(local_settings_dict), so the home settings file overrides repo defaults key-by-key. A ts_adapters list written once and left alone silently loses access to every family whose adapters were added later. It fails closed, quietly, per-family.

2. An uninstalled optional adapter — fatally

Several TS adapters shell out to a separate conda env and repo checkout. KinBot (kinbot_ts.py:253) and AutoTST (autotst_ts.py:216) signal "backend not installed" by raising FileNotFoundError out of execute_incore. That propagated through run_job → spawn_ts_jobs → spawn_post_opt_jobs and terminated the entire ARC run.

The observed cost, in consecutive log lines:

Linear successfully found 7 TS guesses for s0_CH2O2 <=> s1_H2O + s2_CO.
...
ERROR: ARC crashed with <class 'FileNotFoundError'>:
The KinBot python executable was not found. Make sure the kinbot_env exists and
KINBOT_PYTHON is configured...

All 7 guesses, and every converged species job before them, lost to a missing optional dependency — one the user opted into precisely because a multi-adapter search is supposed to be redundant.

The fix

Warn in case 1, with the three facts needed to act: what was configured, what the family admits, and that no guess will be produced. Record-and-continue in case 2, appending to unsuccessful_methods.

Deliberate choices:

• Not raising in case 1 — a user may legitimately skip TS work for some families, and the loop also runs for reactions whose TS was supplied by hand.
• Not auto-adding an adapter — silently changing the user's method set is how case 1 became invisible in the first place.
• FileNotFoundError only, not a bare except — it is the documented "backend missing" signal for both adapters that have one, and a broader catch would mask real defects.

Verification

$ python -m pytest arc/scheduler_test.py -q -k spawn_ts_jobs
3 passed in 6.26s

$ python -m pytest arc/scheduler_test.py -q
11 passed, 35 errors        # baseline on main: 9 passed, 35 errors

The 35 errors are pre-existing on main and unrelated — KeyError: 'server1', the test fixture's template servers dict shadowed by a real ~/.arc/settings.py. Same override mechanism as case 1 above, and worth fixing separately.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 64.36%. Comparing base (d033cbb) to head (21ae360).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #996      +/-   ##
==========================================
- Coverage   64.38%   64.36%   -0.03%     
==========================================
  Files         119      119              
  Lines       39601    39609       +8     
  Branches    10269    10271       +2     
==========================================
- Hits        25499    25495       -4     
- Misses      11121    11130       +9     
- Partials     2981     2984       +3     

Flag	Coverage Δ
functionaltests	64.36% <ø> (-0.03%)	⬇️
unittests	64.36% <ø> (-0.03%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR improves the robustness and debuggability of TS adapter spawning in Scheduler.spawn_ts_jobs() by (1) avoiding whole-run termination when an optional TS backend is missing, and (2) emitting an explicit warning when no TS adapter is eligible for a reaction (instead of failing silently).

Changes:

• Catch FileNotFoundError from TS search job spawning, record the adapter in unsuccessful_methods, and continue to the next adapter.
• Warn when the configured ts_adapters list has no eligible adapter for a known/unknown RMG family, making the “no jobs spawned” condition visible.
• Add regression-style unit tests documenting the two failure modes.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
arc/scheduler.py	Adds graceful handling for missing optional TS adapter backends and warns when no TS adapter is eligible to spawn.
arc/scheduler_test.py	Adds regression tests intended to cover “no eligible adapter” and “missing backend” scenarios.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.